STT: OpenWhispr local listening pipeline -> HAPI things-to-do inbox

[heavygee]

Problem

Operator needs a fast, local STT path for capturing spoken "things to do" without leaving the current focus. Existing voice integrations (ElevenLabs/Gemini/Qwen) are cloud realtime voice assistants -- different latency budget, different privacy posture, different intent.

Proposal

Wire the modified OpenWhispr build (local Whisper STT) as a background listener that:

• Captures audio on a global hotkey (or wake phrase like "i need to" / "i want to" -- see feat(voice): dynamic voice list with preview #2)
• Transcribes locally (Whisper.cpp / faster-whisper)
• Posts the resulting text to a new hub endpoint, e.g. POST /api/inbox/things-to-do { text, source: "stt", capturedAt }
• Hub routes the item into a "things to do" inbox (see feat(voice): dynamic voice list with preview #2)

Out of scope here: the inbox UI itself, the keyword/hotkey UX. Those are #2.

Acceptance

• OpenWhispr instance running locally as a systemd service (or background daemon) and surviving reboot
• One end-to-end test: speak a sentence -> transcript lands in the hub inbox within ~2s
• Hub endpoint accepts authenticated posts and persists to a new inbox_items table or similar
• Documented in docs/tooling/ (operator-local)

Notes

• Pair this with feat(voice): dynamic voice list with preview #2 (inbox UI + hotkey).
• Local STT only -- no cloud round-trip for capture.
• Use the existing voice JWT auth pattern if posting to hub from outside the browser.